[msan] Fix SetShadow for mappings at the end of the application address space

Summary:
On PPC64 if you disable ASLR (or run under gdb) you're likely to see
mmap returning a mapping right at the end of the application address
space region. This caused SetShadow to call MEM_TO_SHADOW() on the
last+1 address in the region, which seems wrong to me; how can
MEM_TO_SHADOW() distinguish this from the first address in the following
region?

Fixed by only calling MEM_TO_SHADOW() once, on the start address.

Reviewers: samsonov, wschmidt, eugenis

Reviewed By: eugenis

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D10735

llvm-svn: 240690